Hadoop中CombineFileInputFormat详解

在MR实践中，会有很多小文件，单个文件产生一个mapper，资源比较浪费，后续没有reduce逻辑的话，会产生很多小文件，文件数量暴涨，对后续的hive job产生影响。

所以需要在mapper中将多个文件合成一个split作为输入，CombineFileInputFormat满足我们的需求。

CombineFileInputFormat 原理(网上牛人总结)：

第一次：将同DN上的所有block生成Split，生成方式：

1．循环nodeToBlocks，获得每个DN上有哪些block

2．循环这些block列表

3．将block从blockToNodes中移除，避免同一个block被包含在多个split中

4．将该block添加到一个有效block的列表中，这个列表主要是保留哪些block已经从blockToNodes中被移除了，方便后面恢复到blockToNodes中

5．向临时变量curSplitSize增加block的大小

6．判断curSplitSize是否已经超过了设置的maxSize

a) 如果超过，执行并添加split信息，并重置curSplitSize和validBlocks

b) 没有超过，继续循环block列表，跳到第2步

7．当前DN上的block列表循环完成，判断剩余的block是否允许被split（剩下的block大小之和是否大于每个DN的最小split大小）

a) 如果允许，执行并添加split信息

b) 如果不被允许，将这些剩余的block归还blockToNodes

8．重置

9．跳到步骤1

// process all nodes and create splits that are local
// to a node.
//创建同一个DN上的split
for (Iterator<Map.Entry<String,
List<OneBlockInfo>>> iter = nodeToBlocks.entrySet().iterator();
iter.hasNext();) {

Map.Entry<String, List<OneBlockInfo>> one = iter.next();
nodes.add(one.getKey());
List<OneBlockInfo> blocksInNode = one.getValue();

// for each block, copy it into validBlocks. Delete it from
// blockToNodes so that the same block does not appear in
// two different splits.
for (OneBlockInfo oneblock : blocksInNode) {
if (blockToNodes.containsKey(oneblock)) {
validBlocks.add(oneblock);
blockToNodes.remove(oneblock);
curSplitSize += oneblock.length;

// if the accumulated split size exceeds the maximum, then
// create this split.
if (maxSize != 0 && curSplitSize >= maxSize) {
// create an input split and add it to the splits array
//创建这些block合并后的split，并将其split添加到split列表中
addCreatedSplit(job, splits, nodes, validBlocks);
//重置
curSplitSize = 0;
validBlocks.clear();
}
}
}
// if there were any blocks left over and their combined size is
// larger than minSplitNode, then combine them into one split.
// Otherwise add them back to the unprocessed pool. It is likely
// that they will be combined with other blocks from the same rack later on.
//其实这里的注释已经说的很清楚，我再按照我的理解说一下
/**
* 这里有几种情况：
* 1、在这个DN上还有没有被split的block，
* 而且这些block的大小大于了在一个DN上的split最小值（没有达到最大值），
* 将把这些block合并成一个split
* 2、剩余的block的大小还是没有达到，将剩余的这些block
* 归还给blockToNodes，等以后统一处理
*/
if (minSizeNode != 0 && curSplitSize >= minSizeNode) {
// create an input split and add it to the splits array
addCreatedSplit(job, splits, nodes, validBlocks);
} else {
for (OneBlockInfo oneblock : validBlocks) {
blockToNodes.put(oneblock, oneblock.hosts);
}
}
validBlocks.clear();
nodes.clear();
curSplitSize = 0;
}

第二次：对不再同一个DN上但是在同一个Rack上的block进行合并（只是之前还剩下的block）

// if blocks in a rack are below the specified minimum size, then keep them
// in 'overflow'. After the processing of all racks is complete, these overflow
// blocks will be combined into splits.
ArrayList<OneBlockInfo> overflowBlocks = new ArrayList<OneBlockInfo>();
ArrayList<String> racks = new ArrayList<String>();

// Process all racks over and over again until there is no more work to do.
//这里处理的就不再是同一个DN上的block
//同一个DN上的已经被处理过了（上面的代码），这里是一些
//还没有被处理的block
while (blockToNodes.size() > 0) {

// Create one split for this rack before moving over to the next rack.
// Come back to this rack after creating a single split for each of the
// remaining racks.
// Process one rack location at a time, Combine all possible blocks that
// reside on this rack as one split. (constrained by minimum and maximum
// split size).

// iterate over all racks
//创建同机架的split
for (Iterator<Map.Entry<String, List<OneBlockInfo>>> iter =
rackToBlocks.entrySet().iterator(); iter.hasNext();) {

Map.Entry<String, List<OneBlockInfo>> one = iter.next();
racks.add(one.getKey());
List<OneBlockInfo> blocks = one.getValue();

// for each block, copy it into validBlocks. Delete it from
// blockToNodes so that the same block does not appear in
// two different splits.
boolean createdSplit = false;
for (OneBlockInfo oneblock : blocks) {
//这里很重要，现在的blockToNodes说明的是还有哪些block没有被split
if (blockToNodes.containsKey(oneblock)) {
validBlocks.add(oneblock);
blockToNodes.remove(oneblock);
curSplitSize += oneblock.length;

// if the accumulated split size exceeds the maximum, then
// create this split.
if (maxSize != 0 && curSplitSize >= maxSize) {
// create an input split and add it to the splits array
addCreatedSplit(job, splits, getHosts(racks), validBlocks);
createdSplit = true;
break;
}
}
}

// if we created a split, then just go to the next rack
if (createdSplit) {
curSplitSize = 0;
validBlocks.clear();
racks.clear();
continue;
}

//还有没有被split的block
//如果这些block的大小大于了同机架的最小split，
//则创建split
//否则，将这些block留到后面处理
if (!validBlocks.isEmpty()) {
if (minSizeRack != 0 && curSplitSize >= minSizeRack) {
// if there is a mimimum size specified, then create a single split
// otherwise, store these blocks into overflow data structure
addCreatedSplit(job, splits, getHosts(racks), validBlocks);
} else {
// There were a few blocks in this rack that remained to be processed.
// Keep them in 'overflow' block list. These will be combined later.
overflowBlocks.addAll(validBlocks);
}
}
curSplitSize = 0;
validBlocks.clear();
racks.clear();
}
}

最后，对于既不在同DN也不在同rack的block进行合并（经过前两步还剩下的block），这里源码就没有什么了，就不再贴了

源码总结：

合并，经过了3个步骤。同DN----》同rack不同DN-----》不同rack

将可以合并的block写到同一个split中
下面是实践代码:

原始文件是70M每个的小文件，有些更小，sequence类型，需要自己实现RecordRead(Text就比较简单)，key是byteWrite类型，现在需要减少文件个数，每个文件的大小接近block的大小。

自定义CombineSequenceFileInputFormat：

package com.hadoop.combineInput;

import java.io.IOException;

import org.apache.hadoop.mapreduce.InputSplit;
import org.apache.hadoop.mapreduce.RecordReader;
import org.apache.hadoop.mapreduce.TaskAttemptContext;
import org.apache.hadoop.mapreduce.lib.input.CombineFileInputFormat;
import org.apache.hadoop.mapreduce.lib.input.CombineFileRecordReader;
import org.apache.hadoop.mapreduce.lib.input.CombineFileSplit;

public class CombineSequenceFileInputFormat<K, V> extends CombineFileInputFormat<K, V> {
@SuppressWarnings({ "unchecked", "rawtypes" })
@Override
public RecordReader<K, V> createRecordReader(InputSplit split, TaskAttemptContext context) throws IOException {
return new CombineFileRecordReader((CombineFileSplit)split, context, CombineSequenceFileRecordReader.class);
}
}

实现 CombineSequenceFileRecordReader

package com.hadoop.combineInput;

import java.io.IOException;

import org.apache.hadoop.mapreduce.InputSplit;
import org.apache.hadoop.mapreduce.RecordReader;
import org.apache.hadoop.mapreduce.TaskAttemptContext;
import org.apache.hadoop.mapreduce.lib.input.CombineFileSplit;
import org.apache.hadoop.mapreduce.lib.input.FileSplit;
import org.apache.hadoop.mapreduce.lib.input.SequenceFileRecordReader;
import org.apache.hadoop.util.ReflectionUtils;

public class CombineSequenceFileRecordReader<K, V> extends RecordReader<K, V> {
private CombineFileSplit split;
private TaskAttemptContext context;
private int index;
private RecordReader<K, V> rr;

@SuppressWarnings("unchecked")
public CombineSequenceFileRecordReader(CombineFileSplit split, TaskAttemptContext context, Integer index) throws IOException, InterruptedException {
this.index = index;
this.split = (CombineFileSplit) split;
this.context = context;

this.rr = ReflectionUtils.newInstance(SequenceFileRecordReader.class, context.getConfiguration());
}

@SuppressWarnings("unchecked")
@Override
public void initialize(InputSplit curSplit, TaskAttemptContext curContext) throws IOException, InterruptedException {
this.split = (CombineFileSplit) curSplit;
this.context = curContext;

if (null == rr) {
rr = ReflectionUtils.newInstance(SequenceFileRecordReader.class, context.getConfiguration());
}

FileSplit fileSplit = new FileSplit(this.split.getPath(index),
this.split.getOffset(index), this.split.getLength(index),
this.split.getLocations());

this.rr.initialize(fileSplit, this.context);
}

@Override
public float getProgress() throws IOException, InterruptedException {
return rr.getProgress();
}

@Override
public void close() throws IOException {
if (null != rr) {
rr.close();
rr = null;
}
}

@Override
public K getCurrentKey()
throws IOException, InterruptedException {
return rr.getCurrentKey();
}

@Override
public V getCurrentValue()
throws IOException, InterruptedException {
return rr.getCurrentValue();
}

@Override
public boolean nextKeyValue() throws IOException, InterruptedException {
return rr.nextKeyValue();
}
}

参考资料：http://sourceforge.net/p/openimaj/code/HEAD/tree/trunk/hadoop/core-hadoop/src/main/java/org/openimaj/hadoop/sequencefile/combine/CombineSequenceFileRecordReader.java

main函数比较简单，这里也贴出来下，方便后续自己记忆：

package com.hadoop.combineInput;

import java.io.IOException;

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.conf.Configured;
import org.apache.hadoop.fs.Path;

import org.apache.hadoop.io.BytesWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapreduce.Job;
import org.apache.hadoop.mapreduce.Mapper;
import org.apache.hadoop.mapreduce.lib.input.FileInputFormat;
import org.apache.hadoop.mapreduce.lib.output.FileOutputFormat;
import org.apache.hadoop.mapreduce.lib.output.SequenceFileOutputFormat;
import org.apache.hadoop.util.Tool;
import org.apache.hadoop.util.ToolRunner;

public class MergeFiles extends Configured implements Tool {
public static class MapClass extends Mapper<BytesWritable, Text, BytesWritable, Text> {

public void map(BytesWritable key, Text value, Context context)
throws IOException, InterruptedException {
context.write(key, value);
}
} // END: MapClass

public int run(String[] args) throws Exception {
Configuration conf = new Configuration();
conf.set("mapred.max.split.size", "157286400");
conf.setBoolean("mapred.output.compress", true);
Job job = new Job(conf);
job.setJobName("MergeFiles");
job.setJarByClass(MergeFiles.class);

job.setMapperClass(MapClass.class);
job.setInputFormatClass(CombineSequenceFileInputFormat.class);
job.setOutputFormatClass(SequenceFileOutputFormat.class);
job.setOutputKeyClass(BytesWritable.class);
job.setOutputValueClass(Text.class);

FileInputFormat.addInputPaths(job, args[0]);
FileOutputFormat.setOutputPath(job, new Path(args[1]));

job.setNumReduceTasks(0);

return job.waitForCompletion(true) ? 0 : 1;
} // END: run

public static void main(String[] args) throws Exception {
int ret = ToolRunner.run(new MergeFiles(), args);
System.exit(ret);
} // END: main
} //

性能测试：70M大小的压缩sequence文件，2000个，转换成是700个压缩sequence文件，平均每个200M(可控)，blocksize=256，耗时2分半到3分钟。

存在问题：

合并后会造成mapper不能本地化，带来mapper的额外开销，需要权衡